[BC Breaking] Make tl.ravel keep element orders by default (#5743)

This doesn't break functional backward compatiblity as the new semantic
is a subset of the what was allowed before but it would break
performance backward compatiblity.
The makes it less error prone.

Author: ThomasRaoux

python/test/unit/language/test_standard.py

@@ -98,6 +98,22 @@ def triton_flip_kernel(out_ptr, x_ptr, N: tl.constexpr):
     torch.testing.assert_close(expect, actual)
 
 
+@pytest.mark.interpreter
+def test_ravel(device):
+
+    @triton.jit
+    def triton_ravel(out_ptr):
+        a = tl.arange(0, 256)
+        a = tl.reshape(a, (32, 8))
+        a = tl.ravel(a)
+        tl.store(out_ptr + tl.arange(0, 256), a)
+
+    out = torch.empty((256, ), device=device, dtype=torch.int32)
+    triton_ravel[(1, )](out)
+
+    assert (out == torch.arange(0, 256, device=device)).all()
+
+
 @pytest.mark.interpreter
 @pytest.mark.parametrize("size_i, size_j, size_g", [[5, 7, 3]])
 def test_swizzle2d(size_i, size_j, size_g, device):

python/triton/language/standard.py

@@ -59,14 +59,14 @@ def softmax(x, ieee_rounding=False):
 
 @core._tensor_member_fn
 @jit
-def ravel(x):
+def ravel(x, can_reorder=False):
     """
     Returns a contiguous flattened view of :code:`x`.
 
     :param x: the input tensor
     :type x: Block
     """
-    return core.reshape(x, [x.numel], can_reorder=True)
+    return core.reshape(x, [x.numel], can_reorder=can_reorder)
 
 
 @jit